How to feed caffe multi label data in HDF5 format?

Question

I want to use caffe with a vector label, not integer. I have checked some answers, and it seems HDF5 is a better way. But then I'm stucked with error like:

accuracy_layer.cpp:34] Check failed: outer_num_ * inner_num_ == bottom[1]->count() (50 vs. 200) Number of labels must match number of predictions; e.g., if label axis == 1 and prediction shape is (N, C, H, W), label count (number of labels) must be N*H*W, with integer values in {0, 1, ..., C-1}.

with HDF5 created as:

f = h5py.File('train.h5', 'w')
f.create_dataset('data', (1200, 128), dtype='f8')
f.create_dataset('label', (1200, 4), dtype='f4')

My network is generated by:

def net(hdf5, batch_size):
    n = caffe.NetSpec()
    n.data, n.label = L.HDF5Data(batch_size=batch_size, source=hdf5, ntop=2)
    n.ip1 = L.InnerProduct(n.data, num_output=50, weight_filler=dict(type='xavier'))
    n.relu1 = L.ReLU(n.ip1, in_place=True)
    n.ip2 = L.InnerProduct(n.relu1, num_output=50, weight_filler=dict(type='xavier'))
    n.relu2 = L.ReLU(n.ip2, in_place=True)
    n.ip3 = L.InnerProduct(n.relu1, num_output=4, weight_filler=dict(type='xavier'))
    n.accuracy = L.Accuracy(n.ip3, n.label)
    n.loss = L.SoftmaxWithLoss(n.ip3, n.label)
    return n.to_proto()

with open(PROJECT_HOME + 'auto_train.prototxt', 'w') as f:
f.write(str(net('/home/romulus/code/project/train.h5list', 50)))

with open(PROJECT_HOME + 'auto_test.prototxt', 'w') as f:
f.write(str(net('/home/romulus/code/project/test.h5list', 20)))

It seems I should increase label number and put things in integer rather than array, but if I do this, caffe complains number of data and label is not equal, then exists.

So, what is the correct format to feed multi label data?

Also, I'm so wondering why no one just simply write the data format how HDF5 maps to caffe blobs?

Solution 1

Answer to this question's title:

The HDF5 file should have two dataset in root, named "data" and "label", respectively. The shape is (data amount, dimension). I'm using only one-dimension data, so I'm not sure what's the order of channel, width, and height. Maybe it does not matter. dtype should be float or double.

A sample code creating train set with h5py is:

import h5py, os
import numpy as np

f = h5py.File('train.h5', 'w')
# 1200 data, each is a 128-dim vector
f.create_dataset('data', (1200, 128), dtype='f8')
# Data's labels, each is a 4-dim vector
f.create_dataset('label', (1200, 4), dtype='f4')

# Fill in something with fixed pattern
# Regularize values to between 0 and 1, or SigmoidCrossEntropyLoss will not work
for i in range(1200):
    a = np.empty(128)
    if i % 4 == 0:
        for j in range(128):
            a[j] = j / 128.0;
        l = [1,0,0,0]
    elif i % 4 == 1:
        for j in range(128):
            a[j] = (128 - j) / 128.0;
        l = [1,0,1,0]
    elif i % 4 == 2:
        for j in range(128):
            a[j] = (j % 6) / 128.0;
        l = [0,1,1,0]
    elif i % 4 == 3:
        for j in range(128):
            a[j] = (j % 4) * 4 / 128.0;
        l = [1,0,1,1]
    f['data'][i] = a
    f['label'][i] = l

f.close()

Also, the accuracy layer is not needed, simply removing it is fine. Next problem is the loss layer. Since SoftmaxWithLoss has only one output (index of the dimension with max value), it can't be used for multi-label problem. Thank to Adian and Shai, I find SigmoidCrossEntropyLoss is good in this case.

Below is the full code, from data creation, training network, and getting test result:

main.py (modified from caffe lanet example)

import os, sys

PROJECT_HOME = '.../project/'
CAFFE_HOME = '.../caffe/'
os.chdir(PROJECT_HOME)

sys.path.insert(0, CAFFE_HOME + 'caffe/python')
import caffe, h5py

from pylab import *
from caffe import layers as L

def net(hdf5, batch_size):
    n = caffe.NetSpec()
    n.data, n.label = L.HDF5Data(batch_size=batch_size, source=hdf5, ntop=2)
    n.ip1 = L.InnerProduct(n.data, num_output=50, weight_filler=dict(type='xavier'))
    n.relu1 = L.ReLU(n.ip1, in_place=True)
    n.ip2 = L.InnerProduct(n.relu1, num_output=50, weight_filler=dict(type='xavier'))
    n.relu2 = L.ReLU(n.ip2, in_place=True)
    n.ip3 = L.InnerProduct(n.relu2, num_output=4, weight_filler=dict(type='xavier'))
    n.loss = L.SigmoidCrossEntropyLoss(n.ip3, n.label)
    return n.to_proto()

with open(PROJECT_HOME + 'auto_train.prototxt', 'w') as f:
    f.write(str(net(PROJECT_HOME + 'train.h5list', 50)))
with open(PROJECT_HOME + 'auto_test.prototxt', 'w') as f:
    f.write(str(net(PROJECT_HOME + 'test.h5list', 20)))

caffe.set_device(0)
caffe.set_mode_gpu()
solver = caffe.SGDSolver(PROJECT_HOME + 'auto_solver.prototxt')

solver.net.forward()
solver.test_nets[0].forward()
solver.step(1)

niter = 200
test_interval = 10
train_loss = zeros(niter)
test_acc = zeros(int(np.ceil(niter * 1.0 / test_interval)))
print len(test_acc)
output = zeros((niter, 8, 4))

# The main solver loop
for it in range(niter):
    solver.step(1)  # SGD by Caffe
    train_loss[it] = solver.net.blobs['loss'].data
    solver.test_nets[0].forward(start='data')
    output[it] = solver.test_nets[0].blobs['ip3'].data[:8]

    if it % test_interval == 0:
        print 'Iteration', it, 'testing...'
        correct = 0
        data = solver.test_nets[0].blobs['ip3'].data
        label = solver.test_nets[0].blobs['label'].data
        for test_it in range(100):
            solver.test_nets[0].forward()
            # Positive values map to label 1, while negative values map to label 0
            for i in range(len(data)):
                for j in range(len(data[i])):
                    if data[i][j] > 0 and label[i][j] == 1:
                        correct += 1
                    elif data[i][j] %lt;= 0 and label[i][j] == 0:
                        correct += 1
        test_acc[int(it / test_interval)] = correct * 1.0 / (len(data) * len(data[0]) * 100)

# Train and test done, outputing convege graph
_, ax1 = subplots()
ax2 = ax1.twinx()
ax1.plot(arange(niter), train_loss)
ax2.plot(test_interval * arange(len(test_acc)), test_acc, 'r')
ax1.set_xlabel('iteration')
ax1.set_ylabel('train loss')
ax2.set_ylabel('test accuracy')
_.savefig('converge.png')

# Check the result of last batch
print solver.test_nets[0].blobs['ip3'].data
print solver.test_nets[0].blobs['label'].data

h5list files simply contain paths of h5 files in each line:

train.h5list

/home/foo/bar/project/train.h5

test.h5list

/home/foo/bar/project/test.h5

and the solver:

auto_solver.prototxt

train_net: "auto_train.prototxt"
test_net: "auto_test.prototxt"
test_iter: 10
test_interval: 20
base_lr: 0.01
momentum: 0.9
weight_decay: 0.0005
lr_policy: "inv"
gamma: 0.0001
power: 0.75
display: 100
max_iter: 10000
snapshot: 5000
snapshot_prefix: "sed"
solver_mode: GPU

Converge graph:

Last batch result:

[[ 35.91593933 -37.46276474 -6.2579031 -6.30313492]
[ 42.69248581 -43.00864792 13.19664764 -3.35134125]
[ -1.36403108 1.38531208 2.77786589 -0.34310576]
[ 2.91686511 -2.88944006 4.34043217 0.32656598]
...
[ 35.91593933 -37.46276474 -6.2579031 -6.30313492]
[ 42.69248581 -43.00864792 13.19664764 -3.35134125]
[ -1.36403108 1.38531208 2.77786589 -0.34310576]
[ 2.91686511 -2.88944006 4.34043217 0.32656598]]

[[ 1. 0. 0. 0.]
[ 1. 0. 1. 0.]
[ 0. 1. 1. 0.]
[ 1. 0. 1. 1.]
...
[ 1. 0. 0. 0.]
[ 1. 0. 1. 0.]
[ 0. 1. 1. 0.]
[ 1. 0. 1. 1.]]

I think this code still has many things to improve. Any suggestion is appreciated.

Solution 2

Your accuracy layer makes no sense.

The way accuracy layer works: in caffe accuracy layer expects two inputs
(i) a predicted probability vector and
(ii) ground-truth corresponding scalar integer label.
The accuracy layer than checks if the probability of the predicted label is indeed the maximal (or within top_k).
Therefore if you have to classify C different classes, your inputs are going to be N-by-C (where N is batch size) input predicted probabilities for N samples belonging to each of the C classes, and N labels.

The way it is defined in your net: You input accuracy layer N-by-4 predictions and N-by-4 labels -- this makes no sense for caffe.

Comments

[Romulus Urakagi Ts'ai]

It seems I misunderstood accuracy layer. But if I delete it, the loss layer returns the same error to me. Maybe I need another Loss layer for vector label? I can't find a list of loss layers available.

[Romulus Urakagi Ts'ai]

I tried EuclideanLoss (without accuracy layer), but it returns massive nan.

[Shai]

@RomulusUrakagiTs'ai is it NaN for the very begining? it might be that the loss is too high causing you gradients to "explode" throwing your training away. Try significantly reducing the loss_weight of the loss layer.

[Romulus Urakagi Ts'ai]

Yes, it's NaN. I'll try that, thank you very much!

[Romulus Urakagi Ts'ai]

I think I make things work, should I write a complete "Answer Your Question", or just comments here and make this the accepted answer?

[Shai]

@RomulusUrakagiTs'ai it depends, if the difference between the two answers is only minor, you can add it as a comment here

[Romulus Urakagi Ts'ai]

Quite things differ, I'll post an answer with full codes.

[R.Falque]

Can you explain how the label is defined, is it a binary system?

[Romulus Urakagi Ts'ai]

Yes, I only tried binary system. ON is 1 and OFF is 0.

[tidy]

What is you caffe version? There is an error for me ImportError: cannot import name layers

[Romulus Urakagi Ts'ai]

I currently don't have the machine, this should be the latest version on Oct 2015.

[Hui Liu]

why do we need to run test_net 100 times when calculating accuracy? Why the result of these 100 runs be different?